Extending k8s API and Lesser known neural networks in Go

Slide 1

Slide 1 text

K8S & LESSER KNOWN NEURAL NETWORKS IN GO GOLANG TAIPEI GATHERING #29

Slide 2

Slide 2 text

ABOUT ME ▸ Masters in Cybernetics, CTU, Prague ▸ Started Kubernetes London Meetup ▸ Containers: https://containerops.org/ ▸ Machine Learning: https://mlexplore.org/ ▸ Twitter: @milosgajdos ▸ Github: milosgajdos83

Slide 3

Slide 3 text

TONIGHT KUBERNETES LONDON MEETUP EXTENDING K8S WITH GO NEURAL NETWORKS IN GO QUIZ!

Slide 4

Slide 4 text

LET’S TALK ABOUT KUBERNETES LONDON

Slide 5

Slide 5 text

KUBERNETES LONDON MEETUP ▸ Started w/ Peter Idah (@peteridah) in Feb 2015 ▸ Currently almost 2k members ▸ Growing strong…still

Slide 6

Slide 6 text

No content

Slide 7

Slide 7 text

LET’S TALK ABOUT KUBERNETES

Slide 8

Slide 8 text

KUBERNETES IN NUTSHELL ▸ Users interact with k8s [only] via API service ▸ Everything [useful] is an API object ▸ Users create them, Control Plane manages them ▸ Control Plane is a “Kernel”, APIs are “standard libs”

Slide 9

Slide 9 text

KUBERNETES IS THE POSIX OF THE CLOUD James Strachan, Redhat

Slide 10

Slide 10 text

EXTENDING KUBERNETES [API]

Slide 11

Slide 11 text

API SERVICE SHIPS AS A COMPILED BINARY

Slide 12

Slide 12 text

NEW K8S RELEASE SHIPS QUARTERLY

Slide 13

Slide 13 text

CONTRIBUTING TO K8S IS NOT EASY!

Slide 14

Slide 14 text

WHAT IF I TOLD YOU…ABOUT CDR?

Slide 15

Slide 15 text

CUSTOM RESOURCE DEFINITION (CRD) ▸ Extend k8s API with custom objects ▸ CDRs allow to use familiar UX: kubectl ▸ On their own just “dumb” storage of structured data ▸ Powerful in combination with CONTROLLERS

Slide 16

Slide 16 text

CDR EXAMPLE

Slide 17

Slide 17 text

CRD IN ACTION

Slide 18

Slide 18 text

CONTROLLERS GIVE YOU CDR SUPERPOWERS…SORT OF

Slide 19

Slide 19 text

CDR CONTROLLER IN ACTION

Slide 20

Slide 20 text

CONTROLLER REGISTRATION

Slide 21

Slide 21 text

CDR CONTROLLERS IN GO ▸ Easy to create, well…sort of :-) ▸ Sample controller:  https://github.com/kubernetes/sample-controller ▸ Dig into the code deeper:   https://goo.gl/d12pQK ▸ Go even deeper: API aggregation ▸ Credit: James Munnelly (@JamesMunnelly )

Slide 22

Slide 22 text

LET’S TALK ABOUT NEURAL NETWORKS

Slide 23

Slide 23 text

WHY GO AND NEURAL NETWORKS ▸ Go is fun! Well, most of the time ;-) ▸ Statically typed language, memory safe ▸ Easy concurrency ▸ Reasonably fast [and still improving] runtime

Slide 24

Slide 24 text

PYTHON [R] + CUDA LEADING THE WAY

Slide 25

Slide 25 text

FPGA+GO=OK FOR SOME WORKLOADS

Slide 26

Slide 26 text

BUT STILL NOWHERE NEAR CUDA GPU MAGIC

Slide 27

Slide 27 text

YOU NEED A GOOD NUMERICAL LIBRARY

Slide 28

Slide 28 text

GO FOR NUMERICAL & SCIENTIFIC COMPUTING ▸ gonum: set of Go libraries for scientiﬁc computing  https://www.gonum.org ▸ Numerical computations and optimization problems:  https://github.com/gonum/gonum ▸ Particularly interesting packages:  mat: matrix operations  optimize: optimisation problems ▸ Plotting graphs:  https://github.com/gonum/plot

Slide 29

Slide 29 text

CREATE MATRIX

Slide 30

Slide 30 text

BASIC MATRIX OPERATIONS

Slide 31

Slide 31 text

TRANSFORM MATRIX OPERATIONS

Slide 32

Slide 32 text

MATRIX VIEW OPERATIONS

Slide 33

Slide 33 text

STATISTICS, COMPARISONS ETC.

Slide 34

Slide 34 text

SOME NOTES ▸ The result of Matrix operation stored in receiver: ▸ Receiver must be initialized! ▸ Receiver get mutated by the operation! ▸ Panics when the indexes are out of bounds! ▸ Some operations return Vector, not Matrix! ▸ You can access raw backing data (ﬂoat64 slices)

Slide 35

Slide 35 text

LESSER KNOWN NEURAL NETS…IN GO

Slide 36

Slide 36 text

SELF ORGANISING MAPS (SOM) ▸ Data clustering ▸ Unsupervised learning ▸ gosom: SOM implementation in Go  https://github.com/milosgajdos83/gosom ▸ http://mlexplore.org/2017/01/13/ self-organizing-maps-in-go/

Slide 37

Slide 37 text

HOPFIELD NETWORKS ▸ Associative memory ▸ Supervised learning ▸ gopfield: Hopfield networks in Go  https://github.com/milosgajdos83/gopfield ▸ http://mlexplore.org/2017/03/12/ hopfield-networks-in-go/

Slide 38

Slide 38 text

TIPS AND TRICKS ▸ Sometimes looping over a raw vector slice is faster than multiplying vectors using mat package functions  AVOID UNNECESSARY ALLOCATIONS ▸ Avoid re-computing CPU “expensive” values (pow, exp…)  CACHE AND RE-USE EXPENSIVE COMPUTATIONS ▸ Go Proﬁler is pretty awesome and it’s your best friend  AVOID PREMATURE OPTIMISATIONS

Slide 39

Slide 39 text

SUMMARY ▸ Extending k8s API is simple and can give you super powers ▸ Neural Networks in Go are fun, but Python + CUDA still faster ▸ Stay awesome and keep hacking in Go!

Slide 40

Slide 40 text

REFERENCES ▸ Gist from this talk  https://gist.github.com/milosgajdos83/ de311c901572644481d4b5dd3830d6cc ▸ Awesome list of Go for data science resources  https://github.com/gopherdata/resources ▸ gorgonnia = Go + CUDA  https://github.com/chewxy/gorgonia ▸ Go + FPGA  https://github.com/ReconﬁgureIO